Texture Feature Extraction And Recognition Of Brain MR Images Based On Multi-scale Resolution Analysis

The diagnosis and treatment of brain disease have been the focus of medical research. Themedical images have been a very important tool of medical diagnosis because they have plenty ofinformation. Multi-scale Resolution Analysis(MRA)joints the time domain and frequency domainto present signal. MRA can decompose signal into multiscale and multi-direction, so it’s beenwidely used in recent years. The classification of images based on feature extraction is theemphasis of Conputer Aided Diagnosis(CAD). There are a lot of common classification methods,while the support vector machine(SVM) shows itself talently among these methods because of itstraining error and generalization ability, and it’s also sutable for small sample learning. In thispaper, the features extracted based on MRA are put into Support Vector Machine(SVM) forclassification. The features which are the optimal representation of the texture images are used inthe feature extraction and classification of brain MR images.The main job are as follows:1.Futher study of four transforms based on MRA and the characteristics of each transform:wavelet transform, the dual-tree complex trhasform, the non-subsampled contourlet transform andbrushlet transform. The features extracted based on the four transforms are selected. Then find theoptimal feature vectors that can present texture feature of images in theory. The features extractedare:the features based on grey level co-occurrence matrix,mean and variance of low frequency andmean,variance and enery of high frequency of wavelet transform,the non-subsampled contourlettransform and the dual-tree complex wavelet transform;mean,variance of the high level phaseangle matrix of the brushlet transform.2.Put the features based on the four kinds of transforms into SVM for classification.Theresult shows that the features based on grey level co-occurrence matrix of low frequency cannotpresent the texture images effectively, therefore this kind of feature is abandoned. And the featuresof wavelet cannot present texture images effectively either. The final features extracted are: meanand variance of both low frequency and high frequency of the dual-tree complex wavelettransform, vector group consist of low frequency mean and variance and high frequency enegy ofthe non-subsampled contourlet transform; the vector group consist of mean, variance of phase angle matrix and enegy of high frequency of brushlet transform.3.Feature extracted based on the former three kinds of transforms of four hundred brain MRimages are put into SVM for classification. The experimental results show that, the accuracy rateof mean, variance and enegy feature vector of the dual-tree complex wavelet transform is83percent.The accuracy rates of the other four vector groups are up to90percent. The feature vectorgroups extracted can distinguish the abnormal images from the normal images effectively.